Regulatory capture is not a scandal. It is a pricing mechanism.
The headline crossed my desk the way most policy headlines do โ unread by the market, upvoted by the timeline. David Sacks, the venture investor now sitting inside the policy apparatus, called the "Pace The Frontier" initiative a case of regulatory capture. The quote circulated. The takes followed. And then the AI-adjacent token basket I track did what it always does when someone hands it a structural signal: nothing. Not a bid. Not a flush. Noise.
That non-reaction is the story. A governance fight over who gets to set the entry bar for frontier AI โ arguably the most consequential capital-allocation decision of the decade โ landed as a tweet and priced as a zero. Everyone is watching the models. Almost nobody is watching the rulebook. The rulebook is the product.
I want to be upfront about something, because it shapes everything below. What I have is a news brief, not a document. Five data points. Four of them are Sacks' own position. No publication date, no text of the initiative, no list of its funders, no response from its authors. That is a thin evidentiary base, and anyone who tells you they can derive a confident conclusion from it is selling you a narrative, not an analysis. So I will mark inference as inference. Where I am extrapolating from my own books, I will say so.
Here is what is actually solid. "Pace The Frontier" โ whatever its final shape โ belongs to the family of proposals that treat frontier model development as a pre-deployment gate rather than a post-hoc liability problem. Evaluate before you train. Report before you scale. Certify before you ship. The intellectual lineage runs back to Stigler's 1971 paper on the theory of economic regulation, which made an uncomfortable point that has never been refuted: regulation is a product, and it is supplied to whoever bids for it. Concentrated interests bid. Diffuse interests free-ride. The public pays the invoice and calls it safety.
Now map that onto AI. The safety camp โ frontier labs with mature legal and security functions, parts of the national security apparatus, a slice of academic AI safety โ wants pre-deployment evaluation. The acceleration camp โ the a16z lineage, the Techno-Optimist Manifesto crowd, and, evidently, Sacks โ wants post-hoc liability and market self-policing. Both camps describe their position as being about safety or about innovation. Both are describing a compliance cost structure.
Here is why a crypto reader should care more about this than about the next ten AI model releases. Crypto has already run this experiment three times. We are the laboratory. The AI fight is the same fight, at a larger balance sheet, with better lawyers.
Start with 2017. I was in high school, and I spent three months manually tracking whale wallets on Etherscan, tagging more than fifty suspicious token launches. What I found was not sophisticated. It was choreography. Eighty percent of those ICOs failed for tokenomics reasons, not technical ones โ supply schedules that could not survive their own unlock calendar. Everyone in that cycle screamed for "regulatory clarity." What arrived instead was the SEC's come-in-and-register doctrine: a door that, by construction, could not be walked through. That is capture in its purest form. A rule that only incumbents can satisfy is not a neutral rule. It is a moat with a legal department.
Fast forward to DeFi summer 2020. I put $5,000 of my own savings across five protocols and lost 30% of it in a flash crash I did not model. The lesson was not about gas fees, which is what my twenty-page internal writeup focused on at the time. The lesson was that "permissionless" systems still have admission control. It is just enforced by liquidity depth rather than by lawyers. Look at the interest rate models on the lending markets โ arbitrary curves dressed as market discovery, parameterized by a handful of governance voters, paid for by everyone else in the spread. Different enforcement mechanism, identical distributional outcome.
Then 2024, the cleanest case. I led three analysts through a fifty-page report on how the spot ETF approvals would move traditional asset flows. We tracked roughly $2 billion of net inflows in the first month and correlated them against equity volatility indices. I presented the finding to institutional clients who insisted crypto was uncorrelated. It wasn't. But the more important finding was structural. The wrapper is the product. Institutions received a compliant vehicle. Retail received the beta. That is regulatory capture executing exactly as designed โ and it was not even controversial, because everyone got a number that went up.
The pattern is stable across all three cycles. Capture does not look like a scandal. It looks like a product launch.
So let me do what the headline did not: run the arithmetic. Compliance is a fixed cost, and fixed costs are barriers to entry. A frontier-class compliance stack is not a form. It is an annual program: pre-deployment evaluations, multiple red-team rounds, external security audits, compute disclosure, provenance infrastructure, and counsel that charges by the hour for the privilege of guessing. For a hyperscaler absorbing $50 billion of capex, that program is a rounding error. For a ten-person lab sitting on a $30 million seed, it is a quarter of the runway, and it is due before revenue exists.
That is the entire economics of capture, written as one inequality: the marginal compliance cost is regressive. It takes a larger share of a smaller balance sheet. You do not need a conspiracy for this outcome. You need a well-intentioned rule and a fixed cost. A threshold is a moat wearing a lab coat.
Which brings me to the part of this debate that almost nobody prices, and the reason I think the Sacks headline is underweighted rather than overrated. Call it threshold drift.
Earlier US frameworks tied regulatory triggers to training compute โ the canonical figure being 10^26 floating-point operations. The design intent was surgical: catch only the frontier, leave everyone else alone. But algorithmic efficiency improves on roughly an annual cadence for a fixed capability level. That means a static compute threshold does not stay on the frontier. It descends the capability ladder. Every twelve months, the same number catches a less advanced model. Today's rule for the largest labs is, within three or four years, a rule for anyone training a competitive open model โ without a single legislator lifting a finger.
The rule does not need to be updated to become more exclusionary. It becomes more exclusionary by standing still. That is a ratchet, and it is why "we will only regulate the frontier" is not a promise. It is a promise with a half-life. Anyone building a business model on the assumption that the bar stays high is implicitly short the drift.
Now, the reflexive crypto trade on all of this is obvious, and I think it is mostly wrong. The story writes itself: if frontier training gets gated by pre-deployment review and compute thresholds, the demand does not vanish, it routes. It routes to decentralized compute markets, inference marketplaces, permissionless training runs, agent economies. The crypto compute stack becomes the offshore financial center of the AI economy.
That is a good story. It is not a flow. Be honest about the ceiling: decentralized training today is a rounding error against hyperscale, and most DePIN compute networks are serving inference and rendering workloads, not frontier pretraining. The revenue link between AI token X and frontier AI capex Y is, in most of the names I have looked at, a blog post. Liquidity is a ghost, not a foundation. It shows up when narrative is cheap and disappears when it isn't, and in a bear market it disappears first from exactly the assets that were priced on a story about a story.
What is real is the compliance market itself. Evaluation, red-teaming, audit, provenance, compute attestation. Regulation creates demand for third-party verification, and that demand is denominated in dollars before it is ever denominated in tokens. Which raises the token question nobody wants: if your protocol sells compliance services, who is the buyer, and why would a regulated entity pay for a critical control in a volatile asset? I have not seen a good answer. I have seen a lot of good decks.
Let me stress-test it instead of selling it. Three scenarios, and none of them are priced.
The base case: "Pace The Frontier" is a voluntary framework, Sacks' critique is rhetoric, and nothing enforceable emerges. Compliance remains private ordering. There is no moat and no trade. AI-crypto stays a narrative asset class and decays with the cycle, which in a bear market means it decays faster than the majors. If you hold AI-adjacent tokens purely on policy-thesis grounds, this scenario is your realized loss.
The ratchet: a threshold-based framework passes in at least one major jurisdiction, drift does its work over three to four years, smaller labs either die or relocate, and the compliance vendor layer consolidates into a handful of firms that audit each other. The moat widens automatically. The positioning here is unglamorous: long real compliance infrastructure, long jurisdiction-arbitrage capacity, short pre-revenue frontier names financed on narrative. In crypto terms, long the pick-and-shovel revenue, short the slide deck.
The race to the bottom: global coordination fails and jurisdictions compete to attract training runs. This is the world Sacks' position implies, and it is genuinely better for aggregate innovation. It is also the world where the first serious incident โ an abuse case, an alignment failure with public consequences โ resets the entire political equilibrium in a single news cycle. In that state, the tail risk is not regulatory. It is reputational, and it lands indiscriminately on anything with "AI" in the ticker.
Which of those three is in the price? My read is none. The market is trading AI tokens on model-release sentiment and trading AI regulation as a non-event. Those are the same mistake.
Now the contrarian turn, because I do not think Sacks is wrong so much as conveniently positioned. Regulation is a fixed cost dressed as a public good โ and "capture" is a label that can be applied in both directions. Sacks is a venture investor who became a policymaker. For the incumbents he has backed, marginal compliance cost is close to zero and the benefit of removing rules is enormous. Criticizing capture is not the same as standing outside it. The question the headline refuses to ask is the only one that matters: who decides what counts as capture, and what is their book?
There is a second, harder point that the crypto audience will like less. Crypto's fixation on AI regulation is a displacement activity. In a bear market, crypto's problem is not Washington. It is that the marginal risk dollar found an alternative. Nine-figure seed rounds in AI labs absorb the same LP capital that used to fund token launches, and they absorb it with better optics and no unlock schedule. AI is not just capturing the rules. It is capturing the speculative base. So the lazy trade โ deregulation is good, therefore buy crypto โ gets the sign wrong. Deregulating AI does not send a dollar to your token. It sends capital further into AI. Smart contracts don't price political risk. Most holders are pricing political vibes and calling it thesis.
So I am watching three things, and none of them is a podcast appearance. Watch the initiative's founding document and its funders โ because the funders tell you who the rule is for. Watch Sacks' actual outputs in his first three to six months: an order, a rescission, a budget line, something with a docket number. Rhetoric is free. Dockets are not. And watch the number. If a compute threshold survives into a final text, that number is the moat, and its drift rate is the most valuable derivative in the market โ unpriced, unlisted, and observable to anyone patient enough to watch it move.
Everything else is commentary. And here is the question I would sit with: if the rulebook is the product, who exactly is your counterparty when you buy the narrative that the rulebook is going away?